Proximal-Mismatch Fine-Tuning / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "config": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "results": [
 19          30.65192222595215,
 20          26.219579696655273,
 21          21.692813873291016,
 22          23.84225082397461
 23        ],
 24        "mean": 25.60164165496826
 25      },
 26      {
 27        "config": {
 28          "lr": 0.003,
 29          "weight_decay": 0.0
 30        },
 31        "results": [
 32          13.522106170654297,
 33          14.461689949035645,
 34          14.807339668273926,
 35          12.651887893676758
 36        ],
 37        "mean": 13.860755920410156
 38      },
 39      {
 40        "config": {
 41          "lr": 0.006,
 42          "weight_decay": 0.0
 43        },
 44        "results": [
 45          10.067484855651855,
 46          12.066851615905762,
 47          9.622394561767578,
 48          10.014313697814941
 49        ],
 50        "mean": 10.442761182785034
 51      },
 52      {
 53        "config": {
 54          "lr": 0.003,
 55          "weight_decay": 0.0001
 56        },
 57        "results": [
 58          13.528329849243164,
 59          14.431048393249512,
 60          14.758162498474121,
 61          12.640395164489746
 62        ],
 63        "mean": 13.839483976364136
 64      }
 65    ],
 66    "full": {
 67      "mean": 10.930935978889465,
 68      "std": 1.3983305692789296,
 69      "per_seed": [
 70        10.067484855651855,
 71        12.066851615905762,
 72        9.622394561767578,
 73        10.014313697814941,
 74        8.787652015686035,
 75        11.810267448425293,
 76        12.032318115234375,
 77        13.046205520629883
 78      ],
 79      "n": 8
 80    }
 81  },
 82  "idea": {
 83    "mean": 10.996043920516968,
 84    "std": 1.169447967118747,
 85    "per_seed": [
 86      9.497220993041992,
 87      12.14220142364502,
 88      11.333624839782715,
 89      9.796889305114746,
 90      9.323980331420898,
 91      11.6688232421875,
 92      11.8473482131958,
 93      12.35826301574707
 94    ],
 95    "n": 8
 96  },
 97  "comparison": {
 98    "delta_mean": 0.06510794162750244,
 99    "idea_wins": 5,
100    "n_pairs": 8,
101    "per_seed_diffs": [
102      -0.5702638626098633,
103      0.07534980773925781,
104      1.7112302780151367,
105      -0.2174243927001953,
106      0.5363283157348633,
107      -0.14144420623779297,
108      -0.18496990203857422,
109      -0.6879425048828125
110    ],
111    "p_value": 0.8863,
112    "mde": 0.6389629457421157,
113    "mde_rel_pct": 5.845455018455167,
114    "verdict": "no measurable effect",
115    "system_worked": false
116  },
117  "mechanism_signature": {
118    "quantity": "mean squared action error to quadratic proximal teacher on held-out states",
119    "predicted": "proximal objective should reduce teacher-action mismatch",
120    "observed_note": "computed from trained models on benchmark-derived perturbations; full values recorded below",
121    "baseline_best_cfg": {
122      "lr": 0.006,
123      "weight_decay": 0.0
124    },
125    "idea_best_cfg": {
126      "lr": 0.006,
127      "weight_decay": 0.0
128    },
129    "confirmed": false
130  },
131  "idea_sweep": [
132    {
133      "config": {
134        "lr": 0.001,
135        "weight_decay": 0.0
136      },
137      "full": {
138        "mean": 25.094110250473022,
139        "std": 4.455265918443395,
140        "per_seed": [
141          27.867267608642578,
142          23.728221893310547,
143          22.022008895874023,
144          23.82248878479004,
145          17.760637283325195,
146          31.66156578063965,
147          22.79692268371582,
148          31.093769073486328
149        ],
150        "n": 8
151      },
152      "mean": 25.094110250473022
153    },
154    {
155      "config": {
156        "lr": 0.003,
157        "weight_decay": 0.0
158      },
159      "full": {
160        "mean": 14.121166110038757,
161        "std": 1.3397338959449183,
162        "per_seed": [
163          13.135297775268555,
164          14.45909309387207,
165          14.704912185668945,
166          12.331259727478027,
167          12.489986419677734,
168          13.940347671508789,
169          15.557445526123047,
170          16.35098648071289
171        ],
172        "n": 8
173      },
174      "mean": 14.121166110038757
175    },
176    {
177      "config": {
178        "lr": 0.006,
179        "weight_decay": 0.0
180      },
181      "full": {
182        "mean": 10.996043920516968,
183        "std": 1.169447967118747,
184        "per_seed": [
185          9.497220993041992,
186          12.14220142364502,
187          11.333624839782715,
188          9.796889305114746,
189          9.323980331420898,
190          11.6688232421875,
191          11.8473482131958,
192          12.35826301574707
193        ],
194        "n": 8
195      },
196      "mean": 10.996043920516968
197    }
198  ],
199  "protocol": {
200    "paired_seeds": [
201      0,
202      1,
203      2,
204      3,
205      4,
206      5,
207      6,
208      7
209    ],
210    "epochs": 18,
211    "batch": 128,
212    "grid_union": [
213      {
214        "lr": 0.001,
215        "weight_decay": 0.0
216      },
217      {
218        "lr": 0.003,
219        "weight_decay": 0.0
220      },
221      {
222        "lr": 0.006,
223        "weight_decay": 0.0
224      },
225      {
226        "lr": 0.003,
227        "weight_decay": 0.0001
228      }
229    ],
230    "structural_match": "tabular: loss/regularization intervention"
231  }
232}